Diagnosing mycosis fungoides using AI powered histologic analysis

Abstract Background Mycosis Fungoides (MF) is the most common type of cutaneous T-cell lymphoma. Unfortunately, the diagnosis of MF is challenging due to its shared histological features with benign inflammatory conditions such as dermatitis, psoriasis, and eczema. Typically, MF diagnosis requires integration of clinical, histopathologic, immunohistochemical, and/or molecular evaluation, which can be time-consuming, subjective, and error prone, leading to significant delays in diagnosis and treatment, negatively impacting patient outcomes. We sought to determine if machine learning approaches may enhance MF diagnostic accuracy using skin biopsy whole slide images (WSI's). Design We employed a self-supervised learning model, Histomorphological Phenotype Learning (HPL), with the BarlowTwins algorithm to autonomously extract distinctive features from skin biopsy hematoxylin and eosin (H&E) stained slide images scanned at 20x magnification and tiled into 224px sections. Our training dataset was comprised of 300 sections from 50 patients diagnosed with early-stage MF, and 300 sections from 50 patients with non-MF inflammatory skin conditions (spongiotic and eczematous dermatitis) sourced from NYU Langone Health (2021–2024). The images were split into 70% training, 10% validation, and 20% testing. External validation was performed using an independent dataset from the University of Minnesota using 50 MF and 50 non-MF cases. Attention heatmaps were applied on the slide level to evaluate the spatial distribution of the highly diagnostic image tiles. Result Our model identified seven enriched, and four depleted image tile clusters in MF vs non-MF cases. In the NYU test set, the model achieved 96% accuracy, 94% specificity, 98% sensitivity, an F1 score of 0.96, and an AUC of 96% for slide-level classification. Application of the model to the second dataset without any additional fine-tuning resulted in a nearly identical performance: 98% accuracy, 96% specificity, 100% sensitivity, an F1 score of 0.98 (precision ≈ 0.96, recall = 1.0), and an AUC of 95.5%. Heatmaps demonstrated distinct patterns of diagnostic patches, with subsets of cases showing predictive clusters predominantly in the dermis, while others were predominantly in the epidermis or epidermal-dermal junction. Conclusion Our studies demonstrate that self-supervised learning can detect features in skin biopsies that can accurately diagnose MF solely based on H&E stained images. Attention heatmaps highlighted three patterns of diagnostic feature enrichment – predominantly epidermal, dermal, or at the epidermal-dermal junction, compatible with biological diversity characterized by distinct patterns of neoplastic cell infiltration. Overall, these studies demonstrate that AI-trained models can accurately diagnose MF. Such models have the potential to increase the speed and accuracy of diagnosis, as well as minimize the use of resource-intensive ancillary testing in MF, leading to improve patient outcomes.